DiffusionRL: Efficient Training of Diffusion Policies for Robotic Grasping Using RL-Adapted Large-Scale Datasets

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Makarova, Maria, Liu, Qian, Tsetserukou, Dzmitry
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911516599517184
author Makarova, Maria
Liu, Qian
Tsetserukou, Dzmitry
author_facet Makarova, Maria
Liu, Qian
Tsetserukou, Dzmitry
contents Diffusion models have been successfully applied in areas such as image, video, and audio generation. Recent works show their promise for sequential decision-making and dexterous manipulation, leveraging their ability to model complex action distributions. However, challenges persist due to the data limitations and scenario-specific adaptation needs. In this paper, we address these challenges by proposing an optimized approach to training diffusion policies using large, pre-built datasets that are enhanced using Reinforcement Learning (RL). Our end-to-end pipeline leverages RL-based enhancement of the DexGraspNet dataset, lightweight diffusion policy training on a dexterous manipulation task for a five-fingered robotic hand, and a pose sampling algorithm for validation. The pipeline achieved a high success rate of 80% for three DexGraspNet objects. By eliminating manual data collection, our approach lowers barriers to adopting diffusion models in robotics, enhancing generalization and robustness for real-world applications.
format Preprint
id arxiv_https___arxiv_org_abs_2505_18876
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle DiffusionRL: Efficient Training of Diffusion Policies for Robotic Grasping Using RL-Adapted Large-Scale Datasets
Makarova, Maria
Liu, Qian
Tsetserukou, Dzmitry
Robotics
Diffusion models have been successfully applied in areas such as image, video, and audio generation. Recent works show their promise for sequential decision-making and dexterous manipulation, leveraging their ability to model complex action distributions. However, challenges persist due to the data limitations and scenario-specific adaptation needs. In this paper, we address these challenges by proposing an optimized approach to training diffusion policies using large, pre-built datasets that are enhanced using Reinforcement Learning (RL). Our end-to-end pipeline leverages RL-based enhancement of the DexGraspNet dataset, lightweight diffusion policy training on a dexterous manipulation task for a five-fingered robotic hand, and a pose sampling algorithm for validation. The pipeline achieved a high success rate of 80% for three DexGraspNet objects. By eliminating manual data collection, our approach lowers barriers to adopting diffusion models in robotics, enhancing generalization and robustness for real-world applications.
title DiffusionRL: Efficient Training of Diffusion Policies for Robotic Grasping Using RL-Adapted Large-Scale Datasets
topic Robotics
url https://arxiv.org/abs/2505.18876